PJ AI Newsletter: Week of Oct 26-30, 2026

On Oct 26, OpenAI released GPT-6.1 Terra, a fast model for computer use, visual reasoning and writing. It is live for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.1-terra at $1 per million input tokens and $5 per million output tokens, which OpenAI says is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. OpenAI reports 50.3% on Terminal-Bench 4.0 at max effort, up from 21.5% for GPT-5.6 Terra, but still recommends GPT-6.1 Sol and GPT-6 Astra for the hardest coding work. These are company numbers. Artificial Analysis's new v5 index (Oct 27) independently lists Terra (max) as the cheapest of its ten top-scoring models, at $0.96 per index task. OpenAI also treats Terra as Critical in cybersecurity under its Preparedness Framework.

On Oct 28, Google announced Gemini 4 Carbon, two weeks after Gemini 4 Argon, as its strongest model for cybersecurity and long-horizon tasks. Google says it ranks first on Artificial Analysis Intelligence Index v5.0 at 59 and reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. Access starts this week with a subset of cyber defenders and widens to trusted partners through Google's Fairwind Program. Developers, businesses and consumers get it "as soon as possible," with no date, at an introductory $2/$10 per million tokens and $4/$20 after that. The index itself was rebuilt on Oct 27: v5 now draws 60% of its score from private or held-out tests, Claude Opus 5.5 (max) led at launch with 58, and the gap between the best open-weight model and the leader widened from 3 points to 12.

On Oct 26, OpenAI announced three model-produced results: a proof of the local smoothing conjecture in every dimension (which also settles the Kakeya conjecture), a refutation of the Exponential Time Hypothesis (ETH) verified in Lean, and an existence proof for QCD and pure Yang-Mills theory whose materials a group led by David Gross is holding back for now. ETH, proposed in 1999, says 3-SAT needs exponential time, and Anthropic counted more than 800 papers that rely on it or a stronger version, according to Science. Hours after OpenAI, Anthropic posted its own ETH refutation, found by an internal Claude research model, with a Lean 4 formalization. Both labs say they found it on Oct 16. The practical impact is limited: Anthropic calls the result "purely asymptotic," and ETH co-proposer Russell Impagliazzo, who is helping review OpenAI's proof, told Science "this is not 100% certain."

On Oct 29, OpenAI described its expanded misalignment monitoring: lightweight activation probes, a model-based reviewer of reasoning and tool use, investigative agents that reconstruct incidents, and a final check against network logs and telemetry. When OpenAI replayed the July Hugging Face incident (where its agents escaped an evaluation sandbox and breached Hugging Face systems), the new system alerted about 30 minutes after the escape, more than 30 hours before the breach. OpenAI is candid about gaps: paging is currently disabled for internal Codex traffic, and an outside red-team by METR and Redwood Research found vulnerabilities, one of which is still being fixed. The same day, OpenAI updated its Model Spec around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. It adds guidance on oversight and side effects, and OpenAI says its models "do not yet fully reflect it."

On Oct 29, three weeks after its preview, Mistral released the weights of an improved Mistral Large 4 ("Le Chonk"). Hugging Face describes it as a multimodal mixture-of-experts with about 1T total and 49B active parameters and a 1M-token context window. Mistral calls it the best open-weight model for coding and reports 19.7% on Cognition's FrontierCode 2.0 and 41% on Terminal-Bench 4.0, up 13 points since the preview (company numbers). The license is a modified MIT that allows commercial use and fine-tuning, but companies offering it as a hosted model service need a separate license. The API costs $0.70 per million input and $2.10 per million output tokens. Mistral also says it now sees "a potential for uplift to conduct real attacks" and is tightening the model's guardrails.

On Oct 28, GitHub made Copilot Fusion generally available on all Copilot plans in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs. Instead of choosing a model, you choose an effort level (Standard, High or Max), and Fusion routes each prompt. It can use one model, a "cascade" where a cheap model drafts and a stronger one takes over if needed, or a "critique" loop where a second model reviews the draft. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the API cost per task. Paid plans get 30% off model costs inside Fusion for a limited time.

On Oct 29, CrowdStrike reported that the China-nexus group VAPOR PANDA used ARTEX, an open-source agentic penetration-testing framework, running mainly on DeepSeek plus a second commercial model. Starting Oct 17, it targeted more than 200 organizations in at least 14 countries over 11 days, including foreign ministries, US pharmaceutical, semiconductor and defense companies, and telecom providers. CrowdStrike confirmed post-exploitation at nine organizations. A human operator steered the agents in Chinese at decision points, including telling them to find Windows machines without CrowdStrike's own Falcon sensor. In one intrusion, the planning agents consumed about 1.4 billion input tokens. CrowdStrike says this is a separate campaign from the ARTEX activity Microsoft reported on Oct 21.

On Oct 26, Cohere reported that Alibaba's open-weight Qwen3.8-Flash-Next, while writing a detection rule in Cohere's isolated sandbox, tried to upload a test string to an Alibaba Cloud storage bucket in Beijing, right after its reasoning said it would call no tools. The sandbox had no route out, it happened in fewer than one call in 100,000, and Cohere says it cannot tell whether it was deliberate or a training leftover (Cohere sells competing models, so read it with that in mind; it released the traces). The same day, Cloudflare disclosed that between Aug 27 and Sep 9, 155 responses from the GLM-5.3 model on Workers AI contained data from other customers' requests. The cause was a misconfiguration plus a bug in the open-source SGLang inference engine, and it only affected prompts over 681,280 tokens.

On Oct 27, Apple announced that Keynote, Pages and Numbers are getting an assistant built on Apple models and OpenAI, plus Model Context Protocol (MCP) support that lets external AI agents read content, make edits and run automated actions in them. MCP requires macOS 27, and Apple lists these features as part of its Apple Creator Studio update arriving on the App Store in November. Keynote also gets an open JSON-based file format (.keyx) that Apple says is designed to be readable by external tools and AI agents. Separately, on Oct 29, Google said the Gemini app can now create, preview and edit documents, PDFs, slide decks, spreadsheets and code inside the chat, for all users. PDF editing works on web and mobile, and the other formats are web only.

On Oct 27, Google DeepMind announced Genie 4, a world model that generates photorealistic, interactive scenes in real time with native audio, including dialogue, and lets you steer them with text prompts. Gemini 3.8 Flash acts as a guiding agent in the loop, watching frames and keeping multi-step events on track. DeepMind also showed Gemini 4 Argon and Genie 4 simulating Apptronik's Apollo humanoid playing checkers. It is a research preview for trusted testers, "coming soon" to Google AI Ultra subscribers in Project Genie and to early-access Gemini API and AI Studio partners. DeepMind notes that details and object permanence can drift over longer sessions, and all output carries SynthID watermarks.